The Dynamics of Returns to Education in Kenyan and Tanzanian
Transcript of The Dynamics of Returns to Education in Kenyan and Tanzanian
1
The Dynamics of Returns to Education in Kenyan and Tanzanian Manufacturing
GPRG-WPS-017
Måns Söderbom, Francis Teal, Anthony Wambugu And Godius
Kahyarara
An ESRC Research Group
Global Poverty Research Group
Website: http://www.gprg.org/
The support of the Economic and Social Research Council (ESRC) is gratefully acknowledged. The work was part of the
programme of the ESRC Global Poverty Research Group.
2
The Dynamics of Returns to Education in Kenyan and
Tanzanian Manufacturing*
Måns Söderbom(1), Francis Teal(1), Anthony Wambugu(2) and Godius Kahyarara(1),(3)
April 2005
Abstract We use micro data on manufacturing employees in Kenya and Tanzania to estimate returns to education and investigate the shape of the earnings function in the period 1993-2001. In Kenya, there have been long run falls in the returns to education while for Tanzania there is evidence of rising returns in the 1990s. The earnings functions are convex for both countries and this result is robust to endogeneity. Convexity may be part of the explanation as to how rapid expansion of education in Africa has generated so little growth if expansion has been concentrated at lower levels of education. Keywords: Returns to education, Africa, Kenya, Tanzania, manufacturing. Word count: 8,873. * We are grateful to Jonathan Temple (Editor), two anonymous referees, Marcel Fafchamps, Geeta Kingdon, Jim Malcomson, Janvier Nkurunziza, Margaret Stevens and participants at a CSAE lunchtime seminar and a Labour Economics seminar at the Department of Economics, Oxford, for several constructive comments on earlier versions of the paper. All errors are our own. Correspondence to Söderbom at [email protected] (1) Centre for the Study of African Economies, Department of Economics, University of Oxford, UK (2) Department of Economics, Kenyatta University, Nariobi, Kenya. (3) Economic and Social Research Foundation, Dar es Salaam, Tanzania.
3
1. Introduction
Returns to education remain of central policy concern in both developed and developing
countries. In developed countries observed rises in returns to education have been imputed to
skill biased technical change (Katz and Autor, 1999). In poorer countries such as those of
Sub-Saharan Africa (SSA) it has been argued that returns may have been falling as a result of
rapid expansion of education. As educational supply grows without a commensurate rise in
demand, the probability of getting a job for any given level of education declines and, among
those with jobs, returns may fall.1 The limited evidence of how wages have changed in Africa
in the recent past suggests they have fallen, possibly substantially. Squire and Suthiwart-
Narueput (1997) document that real minimum wages halved in Kenya between 1970 and
1985, and fell even more in Ghana and Zaire. However it is known that in many countries
minimum wages are not enforced and the numbers are uninformative as to how the returns to
education have been affected.
Policy interest focuses not only on the average return to education but on the dispersion
of returns across education levels. A prominent feature of policy towards education in SSA
has been the priority given to expanding primary education (e.g. World Bank, 1995). Such an
emphasis is seen as being justified, in part, by the finding that returns to education are highest
at lower levels (Psacharopoulos, 1994; Psacharopoulos and Patrinos, 2001). The shape of the
earnings function is a key factor for understanding how policies of education expansion will
impact on incomes. If innovations in educational policy impact primarily on those with high
education costs, and the earnings function is concave, then returns to such reforms will be
relatively high. However, the view that the earnings function is concave in education has
1 Bennell has recently argued as follows: �During the 1960s and 1970s, obtaining a �good� job in the rapidly expanding formal sector of the economy provided a powerful incentive for households to invest in primary education (especially for boys). During the last 20 years, however, wage employment opportunities have contracted sharply in many countries as have formal sector incomes especially in the public sector�, Bennell (2002, p.1186). An influential argument that educational expansion will be self-defeating in that it will simply result in higher qualifications being needed for any given job can be found in Dore (1976). The implications of his argument for Kenya and Tanzania in the 1990s can be found in Toyoda (1997) and Cooksey and Riedmiller (1997).
4
recently been challenged for both developed and developing countries.2 If in fact the earnings
function is convex, so that the marginal returns to education are lowest for the individuals
with the least education, giving priority to investment in primary education may have little
impact on poverty unless the individuals affected by the reforms proceed to higher levels of
education.3
For poor developing countries, such as those in SSA, evidence is very limited as to
what the earnings function looks like (its shape) and how the returns to education have
changed over time.4 This paper considers these issues using comparable repeated cross-
section data on workers in manufacturing firms in Kenya and Tanzania over the 1990s. We
also put the results in a longer term context as excellent data exist for the returns to education
in Kenya and Tanzania in 1980 (Knight and Sabot, 1990). We thus fill a significant
information gap. Further, the different educational policies pursued by Kenya and Tanzania in
the period since independence have been argued to constitute close to a natural experiment
(Knight and Sabot, 1990). In the 1980s, while Kenya allowed a rapid expansion of secondary
education, much of it privately financed, Tanzania severely restricted access to secondary
education and introduced wage polices to reduce differentials. In the 1990s educational and
other policies in Tanzania became much more similar to those in Kenya. A comparative
analysis of these two countries over this period will therefore shed light on some of the
2 In a series of papers Bennell (1996a,b; 2002) has argued that the pattern of the returns to education do not follow that asserted by Psacharopoulos (1994). Bennell�s underlying arguments are consistent with the shape being convex. Direct evidence of convexity in some parts of the domain of the earnings function is provided by Belzil and Hansen (2002) for the U.S. and by Kingdon and Unni (2001) and Duraisamy (2002) for India. 3 Throughout the paper we consider the Mincerian returns to education which do not reflect the private costs, other than foregone wages, or the possible non-wage benefits. It may be that the social returns to primary education are high e.g. in terms of health, but we are unable to investigate this with our data. Further, because primary schooling is a necessary input into postprimary, the prospect of postprimary schooling may raise the primary return above the rate as conventionally measured (Appleton, Hoddinott and Knight, 1996). 4 Much of the available evidence is limited to relatively short periods of time. Krishnan, Sellassie and Dercon (1998) show that educational returns did not change in urban Ethiopia despite labour market reforms instituted in early 1990s. In contrast in Uganda, from 1992 to 1999, returns to education increased markedly, Appleton (2002) and in Ghana from 1987 to 1991 there is evidence of rising returns, Canagarajah and Thomas (1997). Where longer run comparisons have been made there is evidence of falls. In South Africa, Moll (1996) reports that returns to primary education declined from 1960 to 1975. In urban labour markets in Kenya between 1978 and 1995 Appleton, Bigsten and Manda, (1999) report declines in returns to education for workers with secondary education and below.
5
general connections between education policy, education and earnings (e.g. whether the
returns to education change when policies change, and if so, how quickly).
The rest of the paper is organised as follows: Section 2 outlines our empirical
framework; Section 3 discusses the data and shows summary statistics; Section 4 shows OLS
estimates of the earnings functions, and provides a comparative analysis over time and across
age groups; Section 5 shows additional results in which education is treated as an endogenous
variable; and Section 6 provides conclusions.
2. The Earnings Model
Our data begin in 1993 and span seven years for Kenya and eight years for Tanzania. Our
main use of these data is to estimate the earnings-education profile, and to investigate if there
is any evidence of changes in the profile over time or differences across age groups at given
points in time. We write our baseline model of earnings as
( ) iiati sfw υ++= iatxαln (1)
where iw is real earnings of individual i, ix is a vector of worker characteristics excluding
education, atα is a parameter vector, is is the years of education, ( )⋅atf is the earnings-
education profile, iυ is a residual, and a and t denote age group and time, respectively.
Variables included in ix are years of tenure, age and age squared, a dummy variable for
whether the individual is a male or not and a dummy variable for whether the individual lives
in the capital city. It is likely that an important effect of education is to enable individuals to
get high-wage jobs (e.g. managerial positions), or to get into certain high-wage sectors or
firms. Our primary objective in this paper is to estimate the total returns to education, which
may partly reflect such selection effects. We therefore do not include in ix variables that may
be channels through which education affects earnings, e.g. occupation, firm size and sector.5
5 See Fafchamps, Benhassine and Söderbom (2004), for an analysis of how much of the total returns to education in eleven African countries is due to sorting across firms and sorting across occupations within firms. See Söderbom, Teal and Wambugu (2005) for an investigation of the relationship between earnings and firm size in Ghana and Kenya.
6
Key for our purposes is the estimation of the earnings-education profile ( )⋅atf . We
adopt a semiparametric approach modelling ( )⋅atf as a piecewise linear spline function with J
nodes at selected levels of education:
( ) ∑=
−⋅+⋅++=J
jjijatiatatiat sssf
10 }0,max{ θββιμ ,
where jθ denotes the position of the jth node, and tμ and aι are time and age group effects
(intercept shifters). Importantly this approach, which is quite flexible, allows for a non-linear
relationship between education and earnings. The coefficient at0β is interpretable as the slope
of the profile in the first education interval (i.e. for the lowest levels of education), while jatβ
for Jj ,...,2,1= is interpretable as the change in the slope of the profile that results from
moving from the education interval },{ 1 jj θθ − to },{ 1+jj θθ , where 00 =θ . The slope of the
earnings function in the interval },{ 1 jj θθ − , Jj ,...,2,1= , is thus given by ∑ =+
j
m matat ββ10 .
Hence, if 0,...,21 ==== Jctctct βββ , the earnings function is linear. Throughout the analysis
we put in nodes of the earnings-education profile ( )⋅atf at 7, 10 and 12 years of education.
With four segments in the profile there is a reasonable number of observations in each
category. We divide the data into two age groups only, where an individual is considered
�young� if his/her age is less than 30 years and �old� otherwise.
It is widely recognised that using OLS to estimate the returns to education from cross-
section data is potentially problematic. The standard concern in the literature is that education
is an endogenous variable, positively correlated with the earnings residual due to unobserved
ability.6 It is also possible that there is heterogeneity in the returns to education at given levels
of education, and that unobserved ability is correlated with the returns (see e.g. Belzil, 2004).
In either case, OLS estimates of the parameters would be biased. We allow for both forms of
endogeneity in Section 5, using a control function approach.
6 However, the common finding in the empirical literature is that estimated returns rise as a result of treating education as an endogenous variable. We return to this point below.
7
3. Data
We use survey data on employees in the manufacturing sectors in Kenya and Tanzania.7 For
both countries we have four years of data: the Kenyan data cover 1993-1995 and 2000; the
Tanzanian data cover 1993, 1994, 1999 and 2001.8 Four broadly defined manufacturing sub-
sectors were surveyed: food processing, textiles and garments, wood and furniture, and metal-
working including machinery. These sub-sectors comprise the bulk of manufacturing
employment in both countries. Large as well as small firms, including informal ones, were
included in the sample, and each wave of the data contains information on 150-220 firms. In
each firm up to 10 workers were interviewed to provide information on personal
characteristics, characteristics of their jobs and information on earnings and allowances. The
aim was to sample employees representing all types of jobs in the firms, e.g. casual workers,
production workers, supervisors, office clerks and managers. There is a panel dimension at
the firm level, but not at the individual level.9
Table 1 shows summary statistics of the key variables in the analysis. To facilitate
comparison across the two countries, earnings are recorded on a monthly basis and expressed
in constant 1993 US Dollars. The average monthly earnings is USD 75 in Kenya and USD 55
7 There are advantages and disadvantages to focussing solely on individuals in the manufacturing sector and not the whole population. Because of significant private ownership the manufacturing sector provides a good basis for interpreting returns to education as returns to productive skills. In the public sector earnings are determined by a number of factors orthogonal to productive ability, and so the returns to education would have a different interpretation in this sector. Further, focusing on one sector only ensures that changes over time are not driven by changes in the relative sizes of different sectors, across which there may be technological differences. A related point is that, for both Tanzania and Kenya, there is no evidence of significant technological progress over the 1990s (Bigsten, 2002; Pack, 2002). We would therefore argue that for our sample technology is constant. A possible disadvantage of focussing only on the manufacturing sector is that the results may be biased by sample selectivity. We discuss this in Section 5. 8 The first three waves of the Kenyan data, and the first two waves of the Tanzanian data, were collected as part of the World Bank�s Regional Program on Enterprise Development (RPED), while the remaining waves of the data were collected by teams from the Centre for the Study of African Economies, University of Oxford. The survey instruments and the sampling design were very similar both over time and across the two countries, thus providing an excellent basis for comparative analysis. For general information on the surveys, see Söderbom (2001) and Bigsten and Kimuyu (eds.) (2002) for Kenya; and Harding, Kaharaya and Rankin (2002), Harding, Söderbom and Teal (2002), for Tanzania. 9 See Bigsten et al. (2000) for a study of the returns to physical and human capital in five African countries. For a panel data analysis based on the firm level data, see Söderbom and Teal (2004).
8
in Tanzania. The average years of education is 9.1 in Kenya and 8.8 in Tanzania, thus the
large earnings differential across the two countries is not matched by a similarly large
differential in years of education. Figures 1 and 2 illustrate the sample distributions of
earnings (in natural logarithms) and education, distinguishing the two age groups. There is
considerable sample variation in both variables. While there is an obvious differential across
the countries in average earnings, the distributions have similar shapes. For education,
however, the sample distributions differ markedly across the countries. In Tanzania, there is a
spike in the data at 7 years of education, while in the Kenyan sample the distribution features
less kurtosis. This pattern of similar earnings distributions and different education
distributions is interesting. If the aggregate supply of education impacts on the returns to
education, and the differences in supply are not mirrored by differences in aggregate demand
for educated workers, we would expect the earnings-education profiles to differ significantly
across the countries. We now turn to regression analysis to investigate the returns to education
in detail.
4. Earnings Function Estimates
In this section we report OLS estimates of the earnings function parameters, assuming
education to be exogenous. All the parameters are treated as fixed coefficients. If at a given
level of education there is variation in the returns across individuals independent of education,
the estimates are interpretable as averages of individual returns. We allow for endogenous
education and for correlation between education and individual returns in Section 5. We
estimate the earnings equation separately for each age group and time period. We show
results for the two countries in Tables 2 and 3, and to facilitate interpretation we show the
predicted earnings-education profiles in Figures 3a-b. Three main results emerge.
First, for both countries there is strong evidence that earnings are non-linear in
education. For 15 of the 16 regressions reported in Tables 2-3 we can reject linearity, and by
implication constant marginal returns to education, at the 10 per cent level of significance or
lower. Most of the coefficients on the max(.) terms are positive, suggesting that earnings are
9
convex in education. Many of the coefficients on max(0,EDUC-12) are relatively large,
indicating sharp increases in the marginal returns to education after 12 years of education.
This is also apparent in the graphs. Indeed, in Tanzania the main reason why earnings are
non-linear in education is that returns are high at high levels of education, and in six out of
eight cases we can accept at the ten per cent level of significance the hypothesis that the
earnings-education profile is linear between 0 and 12 years of education. In contrast, for
Kenya there is evidence of non-linearities in the earnings profile at lower levels of education,
and at the five per cent level we can reject in five cases out of eight the hypothesis that the
earnings-education profile is linear between 0 and 12 years of education.
Second, for both countries we observe changes in the earnings profiles over time. In
Kenya there is a clear upward intercept shift referring to 1995, which was sustained in 2000
for the old age group but not for the young. The shape of the Kenyan profile, however, looks
quite stable over time, except possibly for 1994. In Tanzania the earnings profiles of 1994,
1999 and 2001 exhibit more pronounced non-linearities than those of 1993. Comparing the
last time period to the first, it appears earnings have become more convex over time for both
age groups.
Third, the data suggest that for both countries earnings profiles differ across the two
age groups. In the range (0,12) years of education the profiles are typically steeper for the old
than for the young age group, especially for Tanzania. In the case of Tanzania it also looks as
though the earnings profile is less convex for the old age group than for the young. In both
countries the returns to education for the young age group are typically quite low before the
tertiary level. We test formally for age group differences and time differences in the earnings-
education relationship below.
In addition to these three results we also note the following: earnings, conditional on
the human capital variables, are higher in Kenya than in Tanzania (this is apparent from the
graphs); the earnings-age profile is, in most cases, inverse u-shaped (of course, these are
within age group profiles); the tenure coefficient is small, typically smaller than 0.01 and
insignificant; the male coefficient is usually positive but only significant at the five per cent
10
level in two out of the 16 regressions; and there is a wage premium to working in the capital
city. We can always reject the hypothesis that the earnings-education profile is flat, and
except for the young age group in Tanzania in 1994 and 1999 we can also reject at the five
per cent level the hypothesis that the earnings-education profile is flat between 0 and 12 years
of education. Thus, returns are certainly low at low levels of education, but in most cases they
are significantly different from zero.
We now analyse in more detail how returns to education have changed over time and
how these returns differ across the two countries and the two age groups. Much of the
comparative work on the returns to education across countries uses a linear specification of
the earnings function (e.g. Trostel et al. 2002), implying that the average equals the marginal
return to education. In our data linearity can typically be firmly rejected, thus marginal returns
will differ from the average. We summarise in Table 4 the marginal returns to education for
1993 and 2000/01 for both countries, distinguishing the old and young age groups. We show
averages of the individual marginal effects for the whole samples as well as a breakdown by
educational level.10 We do the latter by calculating average marginal returns within the
following categories: E2 1�4 years; E3 5-7 years; E4 8-11 years; and E5 12+ years. In Table 5
we use this breakdown to show how the predicted earnings differentials in 1993 and 2000/01
compare to 1980, drawing on the study by Knight and Sabot (1990).11
First consider the averages of the marginal returns across all education levels
combined.12 Throughout, these are higher in 2000/01 than in 1993. In column [5] we report p-
values associated with tests of the hypothesis that there is no difference in the average
marginal returns over time (for both age groups). In Kenya there is no evidence of significant
10 The estimation of the marginal effects is based on the two-sided formula
( ) ( ) ( )( ) εεε 2−−+=′ itatitatitat sfsfsf , where 5.0=ε (i.e. half a year). For observations between
node j and j-1, j > 0, this is equal to ∑ =+
j
m matat ββ10 ; for observations at node j this is the average of
the slopes immediately to the right and left of the node, i.e. atjj
m matat βββ ,110 5.0 +=++∑ .
11 When Knight and Sabot did their survey E2 and E3 corresponded to primary education (standards 1-4 and 5-7, respectively), E4 to secondary, and E5 to tertiary education. In 1985 Kenya reformed its education system to 8 years for primary education, 4 years for secondary, and 4 years for university. 12 These results are similar to those obtained from linear specifications (not reported).
11
differences over the two time periods considered here. In Tanzania average marginal returns
rose from 6 per cent in 1993 to 9 per cent in 2001 for the young, and from 8 to 13 per cent for
the old, and we can reject constant averages at the five per cent level. For both countries,
average returns are always higher in the old age group, but we can accept the null hypothesis
that the averages are the same in the two age groups (column 6).
The breakdown of the marginal returns by education level confirms the magnitude of
the convexity in the returns to education already apparent from Figures 3a-b. For both
countries and both age groups, the marginal return on post-secondary education is higher, in
most cases substantially higher, than the returns before the secondary level. In both countries
and for both age groups the average marginal return on education for those with post-
secondary exceeded 27 per cent in 2000/01. There is evidence of a large and statistically
significant increase in the return to post-secondary education in Tanzania between 1993 and
2001. Thus, the main reason for the increase in the average return in Tanzania documented
above is that the earnings function has become significantly more convex.
Next, we look in detail at total wage differentials associated with various levels of
education over a fixed baseline level. As we are using the same classification of education
variables as that adopted by Knight and Sabot (1990), we can also compare the 1990s with
1980.13 Table 5 shows predicted total earnings premiums (in logarithmic form) attributable to
education, using two different baselines: no education (panel A) and E3 (5-7 years of
education) (panel B). In Kenya, there is evidence of differences in the wage differentials
across age groups at relatively high levels of education. On balance, earnings differentials
attributable to education are larger in the old age group than in the young. When the baseline
13 The sample of workers in Knight and Sabot (1990) was drawn from the wage labour forces in Nairobi and Dar es Salaam. Forty-nine per cent of the sampled individuals in Kenya were manufacturing workers, 58 per cent in Tanzania (see Table A-5 in Knight and Sabot). It is possible that the earnings-education profile differs between manufacturing and non-manufacturing wage employees, in which case the long term comparisons reported in Tables 4 and 5 may partly reflect differences in sector compositions in the samples. We do not have any data that would enable us to assess if this is a problem or not.
12
is E3 (Panel B), there is strong evidence of such age group differentials.14 By year 2000, post-
secondary education is associated with an earnings differentials of 1.08 log points (which
corresponds to a differential in levels of 194 per cent) for an individual in the old age group,
and 0.76 (114 per cent) for an individual in the young age group. There is no evidence of
differences between 1993 and 2000 at any of the levels considered here, but there is a uniform
pattern of wage differentials being smaller in 2000 than in 1980 suggesting that returns in
Kenya have indeed fallen over this period.
In Tanzania we observe in most cases increases in the wage differentials between 1993
and 2001, and with E3 as the baseline we can reject at the five per cent level of significance
the hypothesis that wage differentials associated with post-secondary education are constant
over time. There is also evidence of age group differences amongst individuals with post-
secondary education, with lower differentials amongst the young. Comparing over the longer
term, in all cases except E2 (1-4 years) in Panel A, the Tanzanian wage differentials for 1980
are bracketed by that for the old and the young, suggesting that the shape of the earnings
function of 2001 is not dissimilar to the shape of 1980.
We now investigate whether our findings are sensitive to functional form. Our spline
function is flexible but linearity within each segment is imposed, which may be restrictive.
Further, exactly where the earnings profile has kinks is determined by our own �eye-balling�
of the data. These are potential weaknesses. We now consider the effects of modelling the
earnings-education profile as a third-order polynomial (i.e. a cubic), which sidesteps these
particular problems. To conserve space we do not report the results from these regressions
(they are available on request). Instead we show in Figures 4a-b the estimated earnings
profiles. The shapes of these profiles are similar to their spline function counterparts and there
is clear evidence of convexity. The squared and cubed education effects are jointly significant
at the five per cent level or better in all cases except for the young age group in Kenya in year
2 (tests not reported). We have also investigated how the main findings are affected by
14 Recall from Figure 2 that relatively few workers in the sample have less than 7-8 years of education, and so arguably the estimates in Panel B provide a better reflection of the main trends in the data than
13
including sector controls in the regressions (recall that our data contain individuals working in
four manufacturing sub-sectors in the two countries). The results, which are not reported in
order to conserve space, indicate that there are significant wage differences across sectors in
both countries but the estimated earnings profiles do not change much as a result of including
the sector controls. In particular, the general pattern of convexity of the earnings functions
holds within, as well as between, sectors. These results are available on request from the
authors.
5. Endogenous Education
As is well-known, the OLS estimator will give biased estimates of the returns to education if
education is �endogenous�, i.e. correlated with the residual in the earnings equation. A
common concern in the literature is that education may be positively correlated with
unobserved labour market ability, and that the estimates of the returns to education would be
upward biased as a result.15 For our purposes we are concerned not simply with whether the
average returns to education are biased up, but whether our finding of convexity is also due to
our failure to allow for such endogeneity. Further, our data set consists of employees in the
manufacturing sector, and so self-selection based on unobserved factors into the sector may
give rise to selection bias. In this section we report additional results that should be more
robust to these potential problems than the OLS estimates.
In order to be explicit about the role of unobserved ability, we re-write the earnings
function as
( )( ) ( ) iiiiiati vsfw +++= ληλλ ;ln iatxα , (1')
where iλ is a zero-mean variable denoting unobserved ability, )(⋅η is a continuous function,
iv is a residual orthogonal to all random terms on the right hand side of (1'), and the rest of
those in Panel A. 15 Belzil and Hansen (2002) find a strong positive correlation between unobserved market ability and unobserved taste for schooling, thus leading to substantial upward bias in the OLS estimates of the
14
the notation is as in Section 2. Unobserved ability is thus potentially correlated with both
schooling and the earnings-education profile. Suppose, in reduced form, schooling is given by
iii zs λψ +⋅= , (2)
where iz is a vector of variables (instruments) that are independent of iλ and uncorrelated
with iv . Suppose further that, in the spline function f, we have
] )([)( 3210 atatatiati βββλβλ =atβ ,
i.e. unobserved ability affects the slopes of the different segments of the earnings-education
profile but not the differences in the slopes between segments. The parameter at0β is thus
now explicitly a random coefficient.16 For simplicity, let
( )iatiat λδβλβ += 00 )( , (3)
where at0β is a constant and δ a continuous function with ( ) 00 =δ for notational
convenience.
The conventional form of ability bias arises in this framework when )(⋅η is linear and
increasing in iλ , and iλ does not affect atβ (δ = 0).17 For purposes of exposition, abstract
from time and age group effects and assume the earnings equation is linear in education (i.e.
0321 === βββ ):
( ) iiiii vsw ++⋅+= ληλβ 10ln iαx . (4)
With 01 >η and iλ unobserved, OLS estimates of the return to education will be upward
biased. Further, if the return to education is random and correlated with unobserved ability
(i.e. δ is not constant), i.e.
( )( ) ( ) iiiiii vsw ++⋅++= ληλλδβ 10ln iαx , (5)
returns to schooling. A more common finding in the empirical literature, however, is that estimated returns rise as a result of treating education as an endogenous variable, see e.g. Card (2001). 16 It is straightforward to generalize the model and allow atatat 321 ,, βββ to be correlated with iλ . Doing this with our data generally results in imprecise estimates (sometimes very imprecise). We attribute this to the small sample size. 17 See for instance eq. (7) in Card (2001) under bbi = and 01 =k (using Card�s notation).
15
then this will result in a non-linear association between education and earnings in the data
which is not causal. Notice that if 01 >η and δ is increasing in iλ , i.e. individuals who tend
to get a lot of education tend to have high earnings conditional on education and high returns
to education, then failure to control for this unobserved factor in the estimation will result in a
convex earnings-education profile even though the true functional form is linear.
To allow for a flexible non-linear earnings-education profile whilst controlling for
effects of unobserved ability on earnings and returns to earnings, we adopt a two-stage control
function approach.18 In the first stage we run a regression of the form in (2), thus regressing
education on a set of instruments. Based on this regression we estimate the residual, denoted
iλ� . In the second stage we estimate an earnings function in which iλ� is used as a �control
variable� for ability. Provided standard conditions for identification hold, and provided the
instruments iz are independent of iλ and uncorrelated with iv , this procedure will give
consistent estimates of the parameters of interest.19 As discussed by Card (2001), the control
function approach is more robust than 2SLS when slope parameters potentially co-vary with
the unobserved factors of the model. Further, even if all slope parameters are constant, 2SLS
is likely to result in relative imprecise parameter estimates since the model is non-linear in the
endogenous variable.20 These are the two reasons we prefer this estimator over 2SLS. As is
well known, in the special case where the model is of the form in (4), 2SLS and the control
function estimator are equivalent.
18 For a recent discussion of the control function approach in the context of estimating semiparametric models, see Blundell and Powell (2001). Wooldridge (2002a) considers the control function for estimation of parametric non-linear models (probits and tobits). For an early application of the control function approach in the context of estimating earnings functions, see Garen (1984). 19 Zero covariance between iz and iλ is generally not sufficient for consistency since the model is non-linear in the endogenous explanatory variable, see Blundell and Powell (2001). 20 To apply 2SLS to our model we would have to estimate four first stage regressions, modelling each component of the spline function separately, and then use the predictions instead of the actual values in the second stage. A much richer instrument set would thus be required for 2SLS than for the control function estimator. Unlike the control function approach, however, 2SLS does not require independence between the instruments and the unobserved component of the earnings equation - just zero covariance. Thus, 2SLS is less restrictive than the control function, but identification is likely to be harder to achieve in practice. Essentially, what the independence assumption buys us is that we only need to add a univariate function in the first stage, rather than instrumenting for four variables.
16
We require valid exclusion restrictions, i.e. variables that are correlated with
education and uncorrelated with the earnings residual. In the last wave of the data there is
information on the distance to primary school at the age of six and to secondary school at the
age of twelve, as well as on parents� education and main occupations. These are our potential
instruments for education. Distance to school is a supply side measure of education and it
could therefore be reasonably argued that this variable is correlated with education and not
with ability (Card, 2001). Several recent studies, mostly based on U.S. data, have used similar
information to form instruments for education. Family background variables have been used
as instruments for education in many previous studies, primarily on the grounds that such
variables should have no causal effect on earnings.21
In our discussion so far, we have considered a relatively conventional role of
unobserved ability in potentially leading to bias. In Kenya and Tanzania, unlike more
developed economies, having a job in the wage sector is atypical of outcomes in the labour
market. Such sample selection may lead to bias in the OLS estimates and thus in our
comparison of returns over time. Since we do not have data on individuals outside the
manufacturing sector, we are limited in our ability to control for endogenous sample
selection. In particular, we are unable to use a sample selection model along the lines
proposed by Heckman (1976, 1979), since we cannot estimate a participation equation. Can
the control function address the sample selectivity problem? The answer is yes, provided the
instruments are independent of the error term in the selected sample. One example of a model
in which this will apply is when the job selection model is of the form
⎩⎨⎧ >+
=otherwise 0
0 if 1 1i ii
uI
λ, (6)
21 While these variables may have no direct causal effect on earnings it is still possible that they are invalid instruments, e.g. if the ability of parents is correlated with that of their children or if parents with a lot of education (or with certain jobs) can help their children develop skills that are subsequently rewarded in the labour market. Similarly, if parents with highly able children may choose to live close to a school, the distance variable will not be a valid instrument. Furthermore, it is possible the supply-side instruments are weak instruments for selection into post-secondary education, which could result in bias (see Hahn and Hausman, 2003, for a recent overview of the problems posed by weak instruments). For these reasons, a dose of caution in interpreting the instrumental variable results is recommended.
17
where iI is an indicator variable equal to one if the individual has a manufacturing job and
zero otherwise and iu1 is an unobserved factor which is potentially correlated with iv .
Suppose in the population iu1 follows a standard normal distribution and let iii uuv 21 +⋅= ω ,
where ω is a parameter and iu2 is random term independent of iii zu λ,,,1 ix . In this case the
expected value of iv conditional on selection and ability is given by:
( )iviiii IMzIvE λωσλ == ],,,1[ ix ,
where ( )⋅IM is the inverse Mill�s ratio (see e.g. Wooldridge, 2002b, p.522). Estimating the
earnings function using OLS based on the selected sample would thus generate a combination
of conventional ability bias and selection bias. Because ( )⋅IM is decreasing in ability, the
OLS estimates may in fact be downward biased if the selectivity mechanism is sufficiently
strong and 0>ω , even though education and ability are positively correlated in the
population. In contrast, provided the functional form of the control function is flexible enough
to approximate ( ) ( )ivi IM λωσλη + well, this approach will be consistent. More general cases
of sample selection can be more problematic, however. For instance, if education determines
selection then all instruments in iz will generally be correlated with the earnings residual in
the selected sample, in which case they will not be valid instruments. This will be the case
even if 0=ω , since selection depends both on iz and iλ . Of course, if 0=ω and education
is independent of unobserved ability, the earnings equation can be estimated consistently with
OLS on the selected sample, even if education determines selection.
Results
Table 6 reports control function results using the last wave of data for both countries.22 We
begin by checking that parental characteristics and distance to school indeed have explanatory
power for education, a necessary condition for identification. Based on the first stage
regressions, where education is regressed on all exogenous variables, we test for the joint
18
significance of the coefficients on parental characteristics and distance to school. For all four
specifications, we can safely reject the hypothesis that these coefficients are jointly zero
(EXCRES).
Having established this we next consider our most general earnings model, which
takes the form
( ) iij
jijiiii vsssw ++−⋅+⋅+⋅+= ∑=
)�(}0,max{�ln3
10 ληθβλδβiαx ,
and where we approximate )�( iλδ and )�( iλη by third-order polynomial functions. Based on
the results from this specification we test if the coefficients on the cross terms between )�( iλδ
and education are significant. Within our modelling framework this is interpretable as a test
for whether the returns to education are correlated with unobserved ability. For all four
specifications we can accept the hypothesis that there are no such interaction effects
(RETHET). We therefore move on to specification without interaction terms, which should
give us more efficient estimates (to conserve space, we do not report results for the more
general specification; they are available on request). Table 6 thus shows results based on
models of the form
iij
jijii vssw ++−⋅+⋅+= ∑=
)�(}0,max{ln3
10 ληθββiαx .
Compared to the OLS results, the estimated returns for Kenya tend to increase as a result of
treating education as endogenous. For the young age group the convexity gets more
pronounced as a result of the instrumenting, in the sense that there are large rises in the
coefficients on max(0,EDUC-7) and max(0,EDUC-10). For the old age group, the earnings
profile becomes steeper but the shape is similar to that obtained by OLS. For both age groups
we can reject both linearity of the earnings function and exogeneity at the one per cent level.
These results suggest that OLS underestimates the returns to education. For the Tanzanian
samples the effects of controlling for endogeneity of education are smaller, and we can
22 The reported standard errors have been bootstrapped to take into account the two-step procedure.
19
comfortably accept exogeneity. For the old age group the non-linearities in the earnings
function are highly significant, and for the young age group they are marginally significant.
For both age groups there is a sharp increase in the slope of the earnings function as education
increases from 12 to 13 years. These results are all similar to the OLS results. For both
countries we can always reject that the profile is flat � over the whole range as well as over
the range 0-12 years � and we can reject linearity at the one percent level in three out of fours
cases and at the 10 per cent level in the one remaining case. Similar to the OLS results, there
is mixed evidence for whether linearity in the range 0-12 years can be accepted.
A common result in the empirical literature is that the estimated returns to education
increase as a result of treating education as an endogenous variable. We obtain a similar result
for Kenya, which is quite contrary to the idea that unobserved ability leads to an upward bias
in the estimated return to education. Why might this happen? One possibility is that education
is measured with error, resulting in a downward bias in the OLS estimate. Having been
present at many of the interviews ourselves, it is our impression that respondents were able to
recall years of schooling both with ease and with a relatively high degree of accuracy. We
would therefore be wary of attributing too large a role to reporting errors in this context.
For regression models of earnings that are linear in education, it is common to interpret
the estimated returns to education in models controlling for endogeneity as a local average
treatment effect (LATE). Card (2001) discusses how if there is heterogeneity in the returns to
education, instrumental variable estimates may identify the returns to education for a subset
of individuals with relatively high returns to education. Could a similar explanation hold in
our case? If so, we would expect to find significant correlation between the slope parameter
)(0 iat λβ and the residual from the first-stage regression. As already noted, however, we
cannot reject the null that the interaction effects between the first-stage residual and education
are zero. We would therefore argue that there is no strong support in the data for this
explanation.
20
A third possibility is that sample selectivity plays a role. OLS estimates may in fact be
biased downward if the selectivity mechanism is relatively strong. Under certain assumptions
that have already been discussed, the control function estimator will allow for selectivity. This
could explain why the point estimate rises.
A fourth possible explanation why the control function estimates for Kenya are
relatively high is that the instruments are correlated with the residual of the earnings equation.
Compared to most other studies of returns to education in developing countries we would
argue that our data contain what would seem, a priori, relatively good instruments. Still, it is
clearly possible that in our application the instruments are invalid (see footnote 21) and the
control function estimates inconsistent.
6. Conclusions
There is limited empirical evidence on changes in the returns to education in developing
countries over long periods of time. Our data have enabled us to document changes in the
returns to education in Kenya and Tanzania during the 1990s and also allowed a comparison
with earlier work by Knight and Sabot (1990). It is clear that in the long run the pattern across
the two countries has been very different. Kenya has seen long-run falls in the return to
education at the post primary level and Tanzania has not. Indeed we find for Tanzania an
increase in the return to education in the 1990s.
One of the primary hypotheses advanced to explain the increased returns to education
observed in the U.S. is skill biased technical change. This is unlikely to be the reason for the
changes in the returns to education in Kenya and Tanzania, as the rate of technological
progress in manufacturing has been at best modest in these countries over the last decades
(Bigsten, 2002; Pack, 2002). A more likely explanation can be inferred from the work of
Knight and Sabot (1990), who argue that the high returns in Kenya relative to Tanzania
reflected a willingness to allow market processes to work in Kenya relative to Tanzania. Over
the 1990s Tanzanian polices have become much more similar to those of Kenya, and we have
21
shown that by the end of the 1990s the earnings profiles were quite similar in the two
countries.
The second issue on which we have focused has been the shape of the earnings
function. There is strong evidence that the earnings function is convex for both Kenya and
Tanzania. There are significant differences in the earnings profiles across the age groups. In
Tanzania there is increasing convexity over the 1990s, for Kenya remarkable stability. There
is no evidence from our control function estimates that the finding of convexity is driven by
unobserved ability. In the case of Kenya, where we can reject exogeneity at the five per cent
level for both age groups, the control function estimates are generally larger than their OLS
counterparts. We have discussed four possible explanations: that education is measured with
error; that our instruments impact mostly on students with high returns to education; that
individuals select into the manufacturing sector based on unobserved ability; and that the
instruments are invalid. We would argue that selectivity is a plausible explanation, but we do
not rule out the possibility that the instruments are invalid. While we cannot be sure of the
source, our finding that control function estimates of the returns to education are generally
larger than the OLS estimates is consistent with an increasing number of studies for both
developed and developing countries. From a theoretical point of view, convexity of the
earnings function implies that the cost of education must increase faster than the marginal
return, or everyone would get the maximum level of education. In Africa rapidly increasing
costs seems perfectly plausible. Binding credit constraints, for instance, may well drive up the
shadow cost of obtaining tertiary education.
Finally, one of the micro-macro puzzles in the development literature is why at the
macro level the expansion of education in Africa during the last two decades has generated so
little growth, while at the micro level the average returns to education appear high. With
convexity, these two results could be reconciled if the expansion of education has primarily
occurred on relatively flat segments of the earnings function.
22
References
Appleton, Simon, John Hoddinott, and John Knight (1996). �Primary Education as an Input
into Post-primary Education: A Neglected Benefit,� Oxford Bulletin of Economics and
Statistics 58:1, pp. 211-19.
Appleton, Simon (2002). �Education, Incomes and Poverty in Uganda in the 1990s,� paper
presented at the CSAE conference Understanding Poverty and Growth in Sub-Saharan
Africa at St.Catherine�s College, Oxford, March 18-19, 2002.
Appleton, Simon, Arne Bigsten, and Damiano Kulundu Manda (1999). �Educational
expansion and economic decline. Returns to education in Kenya 1978-1995,� CSAE
Working Paper 99.6. Oxford: Centre for the Study of African Economies, Oxford
University.
Belzil, Christian (2004). �On the Specification of Mincerian Wage Regressions with
Heterogeneity, Non-Linearity, Non-Separability, and Heteroskedasticity,� IZA
Discussion Paper No. 1083. Bonn, Germany: IZA.
Belzil, Christian and Jörgen Hansen (2002). �Unobserved Ability and the Return to
Schooling,� Econometrica 70:5, pp.2075-2091.
Bennell, Paul (1996a). �Rates of Return to Education: Does the Conventional Pattern Prevail
in Sub-Saharan Africa?� World Development 24:1, pp.183-99.
Bennell, Paul (1996b). �Using and Abusing Rate of Return: A Critique of the World Bank�s
1995 Education Sector Review,� International Journal of Educational Development
16:3, pp.235-248.
Bennell, Paul (2002). �Hitting the Target: Doubling Primary School Enrollments in Sub-
Saharan Africa By 2015,� World Development 30:7, pp.1179-1194.
Bigsten, Arne (2002). �History and Policy of Manufacturing in Kenya,� pp.7-30 in A. Bigsten
and P. Kimuyu (eds.) Structure and Performance of Manufacturing in Kenya.
Hampshire, UK: Palgrave.
Bigsten, Arne, Paul Collier, Stefan Dercon, Bernard Gauthier, Marcel Fafchamps, Jan Willem
Gunning, Anders Isaksson, Abena Oduro, Remco Oostendorp, Catherine Pattillo, Måns
23
Söderbom, Francis Teal, Albert Zeufack, and Simon Appleton (2000). �Rates of Return
on Physical and Human Capital in Africa�s Manufacturing Sector,� Economic
Development and Cultural Change 48:4, pp.801-827.
Bigsten Arne and Peter Kimuyu (eds.) (2002). Structure and Performance of Manufacturing
in Kenya. Hampshire, UK: Palgrave.
Blundell, Richard and James L. Powell (2001). �Endogeneity in Nonparametric and
Semiparametric Regression Models,� cemmap working paper CWP09/01. London: The
Institute for Fiscal Studies; Department of Economics, UCL.
Canagarajah, Sudharshan and Saji Thomas (1997). �Ghana�s Labor Market (1987-92),�
Policy Research Working Paper 1752. Washington D.C: The World Bank.
Card, David (2001). �Estimating the Return to Schooling: Progress on Some Persistent
Econometric Problems,� Econometrica 69:5, pp.1127-1160.
Cooksey, Brian and Sibylle Riedmiller (1997). �Tanzanian Education in the Nineties: Beyond
the Diploma Disease,� Assessment in Education: Principles, policy & practice 4:1,
pp.121-135
Dore, Ronald (1976). The Diploma Disease. London: George Allen & Unwin.
Duraisamy, P. (2002). �Changes in Returns to Education in India, 1983-94: By Gender, Age-
Cohort and Location,� Economics of Education Review 21:6, pp. 609-622.
Fafchamps, Marcel, Najy Benhassine and Måns Söderbom, 2004. �Earnings and Human
Capital in African Manufacturing,� mimeo. Oxford: Centre for the Study of African
Economies, Department of Economics, University of Oxford.
Garen, John (1984). �The Returns to Schooling: A Selectivity Bias Approach with a
Continuous Choice Variable,� Econometrica 52:5, pp.1199-1218.
Harding, Alan, Godius Kahyarara, and Neil Rankin (2002). �Firm Growth, Productivity and
Earnings in Tanzanian Manufacturing 1992-1999,� CSAE Report REP/2002-04.
Oxford: Centre for the Study of African Economies, Department of Economics,
University of Oxford.
24
Harding, Alan, Måns Söderbom and Francis Teal (2002). �The Tanzanian Manufacturing
Enterprise Survey 2002,� CSAE Report REP/2002-04. Oxford: Centre for the Study of
African Economies, Department of Economics, University of Oxford.
Hahn, Jinyong and Jerry Hausman (2003). �Weak Instruments: Diagnosis and Cures in
Empirical Economics,� American Economic Review 93:2, pp.118-125.
Heckman, James J. (1976). �The Common Structure of Statistical Models of Truncation,
Sample Selection, and Limited Dependent Variables and a Simple Estimator for Such
Models,� Annals of Economic and Social Measurement 5, pp.475-592.
Heckman, James J. (1979). �Sample Selection Bias as a Specification Error,� Econometrica
47:1, pp.153-161.
Katz, Lawrence F. and David H. Autor (1999) �Changes in the Wage Structure and Earnings
Inequality,� in O. Ashenfelter and D. Card (eds.) Handbook of Labor Economics,
Volume 3A. Amsterdam; New York: North Holland.
Kingdon, Geeta Gandhi and Jeemol Unni (2001). �Education and Women�s Labour Market
Outcomes in India,� Education Economics 9:2, pp.173-194.
Knight, John B. and Richard H. Sabot (1990). Education, Productivity and Inequality: The
East African Natural Experiment. Oxford: Oxford University Press. Published for the
World Bank.
Krishnan, Pramila, Tesfaye Gebre Selassie, and Stefan Dercon (1998). �The Urban Labour
Market During Structural Adjustment: Ethiopia 1990-97,� CSAE Working Paper 98.9.
Oxford: Centre for the Study of African economies, University of Oxford.
Moll, P.G. (1996), The Collapse of Primary Schooling Returns in South Africa 1960-90,�
Oxford Bulletin of Economics and Statistics 58:1, pp.185-210.
Pack, Howard (2002). �Africa�s Declining Productivity: Evidence from Tanzania,� mimeo,
University of Pennsylvania.
Psacharopoulos, George (1994). �Returns to Investment in Education: A Global Update,�
World Development 22:9, pp. 1325-1343.
25
Psacharopoulos, George and Harry Anthony Patrinos (2002). �Returns to Investment in
Education: A Further Update,� World Bank Research Working Paper 2881.
Washington D.C: The World Bank.
Söderbom, Måns (2001). �Constraints and Opportunities in Kenyan Manufacturing: Report
on the Kenyan Manufacturing Enterprise Survey 2000,� CSAE Report REP/2001-03.
Oxford: Centre for the Study of African Economies, Department of Economics,
University of Oxford.
Söderbom, Måns and Francis Teal (2004). �Size and Efficiency in African Manufacturing
Firms: Evidence from Firm-Level Panel Data,� Journal of Development Economics 73,
pp.36-394.
Söderbom, Måns, Francis Teal, and Anthony Wambugu (2005). �Unobserved Heterogeneity
and the Relation between Earnings and Firm Size: Evidence from Two Developing
Countries,� Economics Letters, forthcoming
Squire, Lyn and Sethaput Suthiwart-Narueput (1997). �The Impact of Labor Market
Regulations,� World Bank Economic Review 11:1, pp.119-43.
StataCorp (2003). Stata Statistical Software: Release 8.0. College Station, TX: Stata
Corporation.
Toyoda, Toshio (1997). �Kenya 1975-1995: An Introductory Note on Educational
Expansion,� Assessment in Education: Principles, policy & practice 4:1, pp.87-90.
Trostel, Philip, Ian Walker and Paul Woolley (2002). �Estimates of the Economic Return to
Schooling for 28 Countries,� Labour Economics 9:1, pp.1-16.
Wooldridge, Jeffrey M. (2002a). �Unobserved Heterogeneity and Estimation of Average
Partial Effects,� mimeo. East Lansing, MI: Michigan State University.
Wooldrigde, Jeffrey M. (2002b). Econometric Analysis of Cross-Section and Panel Data.
London, England; Cambridge Massachusetts: The MIT Press.
World Bank (1995). �Priorities and Strategies for Education: A World Bank Sector Review,�
Washington, D.C: The World Bank, Education and Social Policy Department.
26
TABLE 1
SUMMARY STATISTICS
Kenya Tanzania
Mean Standard Dev. Mean Standard Dev.
Earnings(1) 74.6 117.4 54.7 71.2
Years of Education 9.1 2.9 8.8 3.5
Age 33.9 9.1 35.5 10.0
Years of Tenure 7.9 7.2 8.1 7.3
Male Dummy 0.85 0.80
Works in Capital City 0.64 0.44
Old Age group(2) 0.57 0.63
Observations 4039 2738 (1) Monthly earnings expressed in constant 1993 USD. (2) An individual belongs to the old age group if he or she is more than 30 years old.
27
TABLE 2
EARNINGS AND EDUCATION IN KENYAN MANUFACTURING
1993 1994 [1] Young [2] Old [3 Young [4] Old
Age 0.035 0.052 -0.021 0.026 (0.200) (0.039) (0.123) (0.037)
Age squared/100 -0.024 -0.032 0.075 -0.019 (0.387) (0.048) (0.247) (0.042)
Tenure -0.002 0.001 0.021 0.012 (0.011) (0.004) (0.013) (0.005)*
Male -0.001 0.135 -0.021 0.088 (0.082) (0.089) (0.068) (0.095)
Capital city 0.367 0.320 0.287 0.368 (0.071)** (0.063)** (0.068)** (0.071)**
Education 0.027 0.061 0.023 0.038 (0.023) (0.017)** (0.021) (0.019)*
max(0,EDUC-7) -0.041 0.027 0.003 0.006 (0.043) (0.040) (0.042) (0.046)
max(0,EDUC-10) 0.247 0.136 0.060 0.220 (0.093)** (0.085) (0.090) (0.120)+
max(0,EDUC-12) -0.009 0.064 0.083 -0.194 (0.120) (0.093) (0.100) (0.205)
Tests (p-values) Earnings profile flat 0.000 0.000 0.000 0.000 Earnings profile linear 0.000 0.000 0.014 0.011 Education ≤12: Earnings profile flat 0.000 0.000 0.001 0.000 Education ≤12: Earnings profile linear 0.019 0.010 0.679 0.034 R-squared 0.288 0.374 0.182 0.261 Observations 429 675 460 488 The table continues on the next page.
28
TABLE 2 CONTINUED
EARNINGS AND EDUCATION IN KENYAN MANUFACTURING
1995 2000 [1] Young [2] Old [3] Young [4] Old
Age 0.243 0.074 0.040 0.047 (0.113)* (0.031)* (0.116) (0.036)
Age squared/100 -0.429 -0.075 -0.024 -0.035 (0.224)+ (0.036)* (0.231) (0.041)
Tenure 0.023 0.006 0.003 0.008 (0.010)* (0.004) (0.012) (0.004)+
Male 0.049 0.121 0.106 -0.031 (0.079) (0.087) (0.077) (0.080)
Capital city 0.342 0.319 0.251 0.201 (0.066)** (0.073)** (0.062)** (0.092)*
Education 0.007 0.009 0.027 0.014 (0.030) (0.020) (0.067) (0.018)
max(0,EDUC-7) 0.042 0.065 0.018 0.124 (0.054) (0.039)+ (0.097) (0.043)**
max(0,EDUC-10) -0.021 0.008 0.103 -0.034 (0.097) (0.092) (0.104) (0.087)
max(0,EDUC-12) 0.240 0.166 0.116 0.277 (0.097)* (0.126) (0.087) (0.094)**
Tests (p-values) Earnings profile flat 0.000 0.000 0.000 0.000 Earnings profile linear 0.000 0.001 0.000 0.000 Education ≤12: Earnings profile flat 0.048 0.000 0.000 0.000 Education ≤12: Earnings profile linear 0.654 0.073 0.280 0.000 R-squared 0.363 0.265 0.377 0.317 Observations 467 570 371 579 Note: The dependent variable is the log of monthly earnings. Estimation is carried out by regressing the log of monthly earnings on the explanatory variables interacted with eight dummy variables of the form 1{age group, year}. Each column in the table reports the coefficients on the explanatory variables for a given age-year combination. Numbers in parenthesis are the estimated standard errors, which are robust to heteroskedasticity and intra-firm correlation. Significance at the 1%, 5% and 10% level is indicated by * , ** and + respectively. The R-squared is obtained from separate age-year regressions. The numbers of observations refer to each age-year sub-sample.
29
TABLE 3
EARNINGS AND EDUCATION IN TANZANIAN MANUFACTURING
1993 1994 [1] Young [2] Old [3] Young [4] Old
Age 0.035 -0.006 -0.068 0.043 (0.112) (0.034) (0.177) (0.047)
Age squared/100 -0.016 0.020 0.179 -0.036 (0.225) (0.038) (0.362) (0.051)
Tenure 0.004 -0.001 0.008 0.005 (0.014) (0.004) (0.020) (0.008)
Male 0.044 0.186 0.083 0.160 (0.069) (0.082)* (0.092) (0.083)+
Capital city 0.068 0.194 0.279 0.371 (0.082) (0.078)* (0.132)* (0.123)**
Education 0.055 0.034 0.019 0.042 (0.029)+ (0.016)* (0.045) (0.029)
max(0,EDUC-7) -0.012 0.048 0.058 -0.024 (0.053) (0.043) (0.091) (0.064)
max(0,EDUC-10) 0.005 0.076 -0.209 0.216 (0.118) (0.099) (0.189) (0.149)
max(0,EDUC-12) 0.085 -0.076 0.561 -0.122 (0.174) (0.093) (0.279)* (0.142)
Tests (p-values): Earnings profile flat 0.000 0.000 0.003 0.000 Earnings profile linear 0.740 0.006 0.057 0.029 Education ≤12: Earnings profile flat 0.015 0.000 0.422 0.007 Education ≤12: Earnings profile linear 0.970 0.008 0.541 0.153 R-squared 0.140 0.318 0.250 0.318 Observations 304 601 176 268 The table continues on the next page.
30
TABLE 3 CONTINUED
EARNINGS AND EDUCATION IN TANZANIAN MANUFACTURING
1999 2001 [1] Young [2] Old [3] Young [4] Old
Age 0.220 0.060 0.265 0.014 (0.121)+ (0.036)+ (0.152)+ (0.037)
Age squared/100 -0.374 -0.052 -0.460 -0.011 (0.240) (0.040) (0.299) (0.040)
Tenure -0.018 -0.000 0.004 0.004 (0.011)+ (0.005) (0.014) (0.004)
Male 0.063 -0.098 0.270 0.058 (0.080) (0.129) (0.083)** (0.074)
Capital city 0.171 0.119 0.271 0.184 (0.089)+ (0.101) (0.087)** (0.088)*
Education -0.010 0.061 0.053 0.030 (0.014) (0.017)** (0.017)** (0.018)+
max(0,EDUC-7) 0.047 0.031 0.068 0.087 (0.041) (0.035) (0.058) (0.042)*
max(0,EDUC-10) -0.043 -0.016 -0.297 -0.083 (0.137) (0.129) (0.169)+ (0.118)
max(0,EDUC-12) 0.216 0.231 0.469 0.269 (0.170) (0.188) (0.177)** (0.154)+
Tests (p-values): Earnings profile flat 0.002 0.000 0.000 0.000 Earnings profile linear 0.003 0.000 0.012 0.000 Education ≤12: Earnings profile flat 0.570 0.000 0.000 0.000 Education ≤12: Earnings profile linear 0.370 0.630 0.179 0.079 R-squared 0.207 0.417 0.304 0.389 Observations 297 433 227 432 See Table 2 for notes.
31
TABLE 4
MARGINAL RETURNS TO EDUCATION: COMPARATIVE ANALYSIS
[1]
1993, young
[2]
1993, old
[3]
2000/01, young
[4]
2000/01, old
[5]
No difference over time
(p-values)(1)
[6]
No difference across age
groups (p-values)(2)
A. Average marginal returns across all levels of education (3) Kenya 0.11 0.13 0.14 0.16 0.42 0.46 Tanzania 0.06 0.08 0.09 0.13 0.04 0.15 B. Average marginal returns at different levels of education(4) Kenya E2 1�4 years 0.03 0.06 0.03 0.01 0.01 0.30 E3 5-7 years 0.01 0.07 0.03 0.06 0.83 0.04 E4 8-11 years 0.12 0.16 0.09 0.12 0.47 0.42 E5 12+ years 0.23 0.28 0.24 0.33 0.67 0.26 Tanzania E2 1�4 years 0.05 0.03 0.05 0.03 0.99 0.60 E3 5-7 years 0.05 0.06 0.09 0.07 0.51 0.86 E4 8-11 years 0.05 0.11 -0.04 0.09 0.32 0.02 E5 12+ years 0.13 0.08 0.29 0.30 0.01 0.87 (1) Joint test of H0: col. [1] = col. [3] & col. [2] = col. [4]. (2) Joint test of H0: col. [1] = col. [2] & col. [3] = col. [4]. (3) Sample averages of the marginal effects of education, evaluated for each individual in the sample. The estimates are based on the regressions reported in Tables 2 and 3. (4) The estimates are based on the regressions reported in Tables 2 and 3.
32
TABLE 5
PREDICTED EARNINGS DIFFERENTIALS AND EDUCATION: COMPARATIVE ANALYSIS
[1]
1993, young
[2]
1993, old
[3]
2000/01, young
[4]
2000/01, old
[5]
No difference over time
(p-values)(1)
[6]
No difference across age
groups (p-values)(2)
[7]
1980(3)
A. Earnings differentials. Baseline: No education(4) Kenya E2 1�4 years 0.09 0.20 0.09 0.05 0.15 0.47 0.18 E3 5-7 years 0.18 0.41 0.18 0.09 0.15 0.47 0.38 E4 8-11 years 0.15 0.68 0.31 0.49 0.52 0.03 0.86 E5 12+ years 0.88 1.49 0.94 1.18 0.23 0.01 1.62 Tanzania E2 1�4 years 0.20 0.13 0.19 0.11 0.99 0.56 0.06 E3 5-7 years 0.38 0.24 0.37 0.21 0.99 0.56 0.28 E4 8-11 years 0.50 0.46 0.70 0.53 0.72 0.74 0.64 E5 12+ years 0.88 0.96 0.97 1.24 0.32 0.35 1.04 B. Earnings differentials. Baseline: E3 5-7 years (5) Kenya E4 8-11 years -0.03 0.27 0.14 0.39 0.41 0.01 0.48 E5 12+ years 0.70 1.08 0.76 1.08 0.90 0.00 1.24 Tanzania E4 8-11 years 0.12 0.22 0.33 0.32 0.40 0.74 0.36 E5 12+ years 0.50 0.73 0.60 1.03 0.05 0.00 0.76 (1) Joint test of H0: col. [1] = col. [3] & col. [2] = col. [4]. (2) Joint test of H0: col. [1] = col. [2] & col. [3] = col. [4]. (3) These estimates are based on Knight and Sabot (1990), Table 6-2, column 3. (4) Predicted log earnings differentials between given levels of education and the baseline level, which is no education. For example, the predicted earnings of a Kenyan individual in the young age group in 1993 with 1-4 years of education is 0.09 log points higher than an otherwise identical individual without education. This corresponds to a differential of about exp(.09) - 1 = 9 per cent. The estimates are based on the regressions reported in Tables 2 and 3. (5) As panel B except that the baseline is E3 5-7 years of education. The estimates are based on the regressions reported in Tables 2 and 3.
33
TABLE 6
CONTROL FUNCTION ESTIMATES: KENYA 2000 AND TANZANIA 2001
Kenya Tanzania [1] Young [2] Old [3] Young [4] Old Age -0.031 0.029 0.289 0.008 (0.127) (0.047) (0.142)* (0.038)
Age squared/100 0.062 -0.004 -0.532 -0.004 (0.250) (0.054) (0.281)+ (0.043)
Tenure 0.026 0.013 0.012 0.005 (0.014)+ (0.005)** (0.017) (0.004)
Male 0.270 0.005 0.352 0.068 (0.094)** (0.080) (0.086)** (0.080)
Capital city 0.234 0.207 0.248 0.183 (0.060)** (0.093)* (0.088)** (0.091)*
Education -0.060 0.113 0.106 0.022 (0.169) (0.052)* (0.063)+ (0.038)
max(0,EDUC-7) 0.202 0.109 0.050 0.115 (0.182) (0.060)+ (0.086) (0.049)*
max(0,EDUC-10) 0.154 -0.035 -0.280 -0.081 (0.098) (0.086) (0.189) (0.121)
max(0,EDUC-12) 0.099 0.313 0.463 0.258 (0.099) (0.098)** (0.196)* (0.181)
Tests (p-values) Earnings profile flat 0.00 0.00 0.00 0.00 Earnings profile linear 0.00 0.00 0.06 0.00 Education ≤12: Earnings profile flat 0.00 0.00 0.03 0.00 Education ≤12: Earnings profile linear 0.05 0.12 0.31 0.04 EXCRES(1) 0.00 0.00 0.00 0.00 RETHET (2) 0.16 0.60 0.95 0.61 EXOGEN(3) 0.00 0.00 0.35 0.58
Observations 371 579 227 432
Note: The dependent variable the log of monthly earnings. Education is endogenous. Bootstrapped standard errors, robust to heteroskedasticity and intra-firm correlation are reported in parenthesis. Significance at the 1%, 5% and 10% level is indicated by * , ** and + respectively. The explanatory variables in the regression modelling education (step 1) are: dummy variables for mother's and father's education (none; primary; middle (pre 1964) or training college; O level; A level; vocational/technical; university; don't know) and occupation (farming, fishing, forestry; trading, self-employed; clerical; employed in construction, tailoring, or worked as foreman; professional; watchman, soldier; don�t know); dummy variables for distance to primary school at the age of six and to secondary school at the age of twelve (less than 1 km; 1-3 km; 3-6 km; 6-10 km; more than 10 km; don�t know); and age, age squared, tenure, male, capital city. (1) Wald test for joint significance of the identifying instruments in step 1 (the exclusion restrictions). (2) Wald test for joint significance of the interaction terms between education and a third degree polynomial in the first stage residual. (3) Wald test for joint significance of a third degree polynomial in the first stage residual.
34
FIGURE 1
SAMPLE DISTRIBUTIONS OF LOG EARNINGS
0
0.2
0.4
0.6
0.8
1.5 2.5 3.5 4.5 5.5 6.5 7.5
ln Earnings
Kenya, young Kenya, old Tanzania, young Tanzania, old
Note: The density estimates were obtained using the Stata 8.0 command �kdensity� (StataCorp, 2003).
35
FIGURE 2
SAMPLE DISTRIBUTIONS OF YEARS OF EDUCATION
i) Kenya
0
0.1
0.2
0.3
0.4
0.5
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
Years of education
Sam
ple
prop
ortio
n
Young Old
ii) Tanzania
0
0.1
0.2
0.3
0.4
0.5
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
Years of education
Sam
ple
prop
ortio
n
Young Old
36
FIGURE 3A
EARNINGS AND EDUCATION: KENYA i) Age group: Young
2.5
3
3.5
4
4.5
5
5.5
6
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
Education (years)
Pred
icte
d lo
g ea
rnin
gs
Year 1 Year 2 Year 3 Year 4
ii) Age group: Old
2.5
3
3.5
4
4.5
5
5.5
6
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
Education (years)
Pred
icte
d lo
g ea
rnin
gs
Year 1 Year 2 Year 3 Year 4
Note: Year 1 = 1993; Year 2 = 1994; Year 3 = 1995; Year 4 = 2000.
37
FIGURE 3B
EARNINGS AND EDUCATION: TANZANIA i) Age group: Young
2.5
3
3.5
4
4.5
5
5.5
6
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
Education (years)
Pred
icte
d lo
g ea
rnin
gs
Year 1 Year 2 Year 3 Year 4
ii) Age group: Old
2.5
3
3.5
4
4.5
5
5.5
6
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
Education (years)
Pred
icte
d lo
g ea
rnin
gs
Year 1 Year 2 Year 3 Year 4
Note: Year 1 = 1993; Year 2 = 1994; Year 3 = 1999; Year 4 = 2001.
38
FIGURE 4A
EARNINGS AND EDUCATION, POLYNOMIAL SPECIFICATION: KENYA i) Age group: Young
2.5
3
3.5
4
4.5
5
5.5
6
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
Education (years)
Pred
icte
d lo
g ea
rnin
gs
Year 1 Year 2 Year 3 Year 4
ii) Age group: Old
2.5
3
3.5
4
4.5
5
5.5
6
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
Education (years)
Pred
icte
d lo
g ea
rnin
gs
Year 1 Year 2 Year 3 Year 4
Note: Year 1 = 1993; Year 2 = 1994; Year 3 = 1995; Year 4 = 2000.
39
FIGURE 4B
EARNINGS AND EDUCATION, POLYNOMIAL SPECIFICATION: TANZANIA i) Age group: Young
2.5
3
3.5
4
4.5
5
5.5
6
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
Education (years)
Pred
icte
d lo
g ea
rnin
gs
Year 1 Year 2 Year 3 Year 4
ii) Age group: Old
2.5
3
3.5
4
4.5
5
5.5
6
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
Education (years)
Pred
icte
d lo
g ea
rnin
gs
Year 1 Year 2 Year 3 Year 4
Note: Year 1 = 1993; Year 2 = 1994; Year 3 = 1999; Year 4 = 2001.